fix: chunk-and-summarize execution, gemini-3.5-flash window, and Phase 3/4 checklist discrepancies #8

Problem Description

A comprehensive review of Phases 3 & 4 against the consolidated requirements and review checklist identified several discrepancies, missing deliverables, and bugs:

1. Missing Chunk-and-Summarize Execution for Large Diffs:
   When TokenBudgetManager.assess() returns BudgetStrategy.NEEDS_CHUNKING, generate_commit_message in src/gitmate/fallback.py previously short-circuited directly to a template-based fallback (is_fallback=True), skipping LLM generation entirely. As specified in plan.md and the Phase 4 review checklist, Phase 4 must consume BudgetDecision.omitted_diffs (which preserves the untruncated original patches), summarize each chunk via active_provider.generate(), merge the summaries into the prompt's summary_note, and generate the final commit message (with graceful degradation if the provider fails).

2. Missing gemini-3.5-flash in DEFAULT_MODEL_CONTEXT_WINDOWS:
   DEFAULT_MODEL_CONTEXT_WINDOWS in src/gitmate/config.py contained gemini-3.5-flash-lite, gemini-3-flash, and gemini-flash, but omitted gemini-3.5-flash (non-lite) and lacked an explanatory comment clarifying that gemini-3.5-flash-lite remains the configured default. Users setting model = "gemini-3.5-flash" received UnknownModelError.

3. Missing RetryableError Alias in providers/base.py:
   The transient error marker was named RetryableProviderError, but the review checklist and callers expect RetryableError.

4. Stale SDK and Provider References in plan.md:
   plan.md retained stale references to AnthropicProvider (Phase 4, line 90) and Anthropic SDK token counting (Phase 2, lines 57 & 63, and design decision table line 15), contradicting the locked-in decision in AGENTS.md specifying Gemini as default.

5. template_overhead Calibration:
   The 500 token default estimate was evaluated against the actual template files (commit_conventional.txt: ~329 tokens, commit_plain.txt: ~210 tokens, pr_summary.txt: ~206 tokens), confirming 500 tokens provides a safe upper bound with ~170+ tokens of headroom for dynamic notes.

Proposed Resolution

• Implement _chunk_patch_text and summarize_omitted_diffs in src/gitmate/fallback.py to execute chunk-and-summarize on NEEDS_CHUNKING, merging notes into the final prompt while retaining graceful degradation on provider failure.
• Add "gemini-3.5-flash": 1_048_576 to DEFAULT_MODEL_CONTEXT_WINDOWS in src/gitmate/config.py with explanatory comments.
• Export RetryableError = RetryableProviderError in src/gitmate/providers/base.py.
• Update plan.md to eliminate stale Anthropic SDK references in favor of the Google GenAI SDK and document Phase 4 chunk-and-summarize delivery.
• Add test coverage across tests/test_orchestrator.py, tests/test_config.py, and tests/test_providers.py.